Download CONVEXITY AND HAAR NULL SETS A Borel subset A of a

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts
no text concepts found
Transcript
PROCEEDINGS OF THE
AMERICAN MATHEMATICAL SOCIETY
Volume 125, Number 6, June 1997, Pages 1793–1799
S 0002-9939(97)03776-3
CONVEXITY AND HAAR NULL SETS
EVA MATOUŠKOVÁ
(Communicated by Dale Alspach)
Abstract. It is shown that for every closed, convex and nowhere dense subset
C of a superreflexive Banach space X there exists a Radon probability measure
µ on X so that µ(C + x) = 0 for all x ∈ X. In particular, closed, convex,
nowhere dense sets in separable superreflexive Banach spaces are Haar null.
This is unlike the situation in separable nonreflexive Banach spaces, where
there always exists a closed convex nowhere dense set which is not Haar null.
A Borel subset A of a separable Banach space X is called a Haar null set if
there exists a probability measure µ on the σ-algebra of Borel subsets of X so that
µ(A+x) = 0 for all x ∈ X (see [C] also for the following properties of Haar null sets).
The family of all such sets is closed under translation and under countable unions;
nonempty open sets do not belong to it. The Haar null sets agree with Lebesgue
null Borel sets in finite dimensional spaces. This definition of null sets is rather
weak, as every compact set in an infinite dimensional space is a Haar null set. In
fact, in infinite dimensional superreflexive and nonreflexive Banach spaces even all
weakly compact convex sets with empty interior are Haar null. For superreflexive
spaces this follows from our result. S
If K is a weakly compact and convex subset of
a nonreflexive Banach space, then t>0 t(K − K) 6= X and there exists x ∈ X so
that the intersection of the line segment [0, x] and any translate of K contains at
most one point. Consequently, if µ is Lebesgue measure on [0, x], then µ(K +x) = 0
for each x ∈ X. In [MS] it is shown that a separable Banach space is nonreflexive
if and only if there exists a closed convex subset Q of X with empty interior, which
contains a translate of any compact subset of X. Such a set Q is not Haar null
because given any probability measure µ on X there exists a compact set K with
µ(K) > 0 and, consequently, also a translate of Q of positive measure. It follows
that every separable nonreflexive Banach space contains a closed convex set with
empty interior which is not Haar null. In this note we show that this is unlike the
situation in superreflexive spaces, where for every closed convex set C with empty
interior there exists a Radon probability measure µ on X so that µ(C + x) = 0
for all x ∈ X. We do not know if such a measure exists when X is only reflexive
or, equivalently, if the positive cone of every reflexive Banach space with a basis is
Haar null.
Received by the editors February 22, 1995 and, in revised form, January 8, 1996.
1991 Mathematics Subject Classification. Primary 46B10; Secondary 46B20.
Key words and phrases. Superreflexive Banach spaces, convexity, Haar null sets.
The author was partially supported by the grant GAČR 201/94/0069 and by a grant of the
Austrian Ministry of Education.
c
1997
American Mathematical Society
1793
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
1794
EVA MATOUŠKOVÁ
There are several results dealing with the question of how often a convex continuous function on a Banach space is differentiable. Let us mention a few examples
of the “almost everywhere” ones. In the finite dimensional spaces the theorem of
Rademacher says that every locally Lipschitz (and, hence, every convex and continuous) function is almost everywhere Fréchet differentiable. In separable Banach
spaces the substitute for this theorem is Aronszajn’s theorem [A], which says that
every locally Lipschitz function is Gâteaux differentiable outside of an Aronszajn
null set. This does not hold for Fréchet differentiability even in reflexive spaces as
the Hilbert cube D = {x ∈ `2 : 0 ≤ xn ≤ 1/n} is not Aronszajn null (cf. [A])
and the distance function to D (which is convex and continuous) is not Fréchet
differentiable at any point of D. One may wonder if it would help to replace the
Aronszajn null sets by some larger class of null sets, for example, the Haar null sets.
This is not the case for functions which are just Lipschitz; from [PT] it follows that
on every separable Banach space there exists a Lipschitz function for which the set
of points where it is not Fréchet differentiable is not Haar null. In every separable
nonreflexive space, there even exists a convex continuous function with this property. Indeed, the distance function to a closed convex subset K of a Banach space
is not Fréchet differentiable at any point of the boundary of K, so the distance
function to a closed convex set which is not Haar null and has empty interior will
do as an example. We do not know whether such an example exists in separable
reflexive spaces. From our result, which answers negatively a conjecture in [BN], it
follows that in separable superreflexive space one can not construct an example as
above since the boundaries of convex sets are Haar null. For sets with nonempty
interior this is true in any separable Banach space. Indeed, it is easy to see that if
K is a closed convex set with the origin in the interior then the supremum of the
Minkowski functional of K and the constant function equal to one is not Gâteaux
differentiable at any point of the boundary of K. Consequently, the boundary is,
by Aronszajn’s theorem, Aronszajn null and hence also Haar null. If the interior
of K is empty, this trick does not always work. If K is not Aronszajn null (for example, take again the Hilbert cube D in `2 as above) then every convex continuous
function is Gâteaux differentiable also at some points of K. Finally, let us remark,
that if instead of Fréchet smoothness we consider the stronger notion of Lipschitz
smoothness then there exists a convex continuous function on `2 such that the set
of points where it is not smooth is not Haar null [MZ].
Let X be a Banach space with a Schauder basis {xn }. We say that the basis
is normalized, if kxn k = 1 for all n ∈ N .PBy the positive cone of X we mean the
∞
i = 1, 2, . . . }.
(necessarily closed and convex) set Q = { i=1 αi xi ∈ X : αi ≥ 0 for P
n
The characteristic of the basis is the largest of all ε ≥ 0 for which k i=1 ai xi k ≥
Pk
εk i=1 ai xi k if n > k and {ai } is any sequence of numbers; it holds always that
0 < ε ≤ 1. A Radon measure on a Banach space X is a measure on the σ-algebra
of Borel subsets of X, which is inner regular with respect to the compact sets. We
say that a convex subset C of a Banach space X is spanning if it contains a line
segment in every direction, that is,
[
t(C − C) = X.
t>0
In [MS] it is shown that if Q is a closed, convex, bounded, spanning and nowhere
dense subset of a Banach space X, and T : X → Y is a surjective, continuous and
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
HAAR NULL SETS
1795
linear operator with finite dimensional kernel, then T (Q) is also a closed, convex,
bounded and spanning subset of Y with empty interior.
Lemma 1. Let X be a weakly compactly generated Banach space, and C ⊂ X a
closed, convex, bounded and spanning set with empty interior. Then there exists
an infinite dimensional quotient space Y of X with a Schauder basis and v ∈ X
so that the image of C + v under the quotient mapping is contained in the positive
cone of Y .
Proof. If C is a closed, convex and nowhere dense subset of a weakly compactly
generated Banach space X, then there exists a projection P onto a separable subsace
Z of X such that the interior of P (C) in Z is empty (see Lemma 3). Therefore,
we may suppose that X is separable. Let us recall the following elementary fact: if
K is a closed, convex and nowhere dense subset of a Banach space then for every
ε > 0 there exists a norm one element x∗ of the dual so that hx∗ , xi > −ε for all
x ∈ K. Choose a sequence {xn } that is dense in the unit sphere of X. For n ≥ 1
construct, by induction, a sequence {x∗n } in the unit sphere of X ∗ and a sequence
{vn } in X as follows. Set k0 = 1. Put Xn = span{x1 , . . . , xkn−1 } and choose a
closed subspace Zn of X so that Xn ⊕ Zn = X. Let Pn : X → Zn be the projection
with kernel Xn . Then Pn∗ is an isomorphism between Zn∗ and the annihilator Xn⊥
of Xn in X ∗ ; choose cn > 0 so that cn kz ∗ k ≤ kPn∗ (z ∗ )k for each z ∗ ∈ Zn∗ . Let Cn
be the projection of C in Zn . Since C is closed, convex, bounded, nowhere dense
and spanning, Cn is a closed convex set with empty interior by the result from [MS]
mentioned above. Consequently, there exists zn∗ ∈ Zn∗ so that
kzn∗ k = 1
and hzn∗ , zi ≥ −cn /4n+1
for each z ∈ Cn . Define x∗n = Pn∗ (zn∗ )/kPn∗ (zn∗ )k. Then x∗n ∈ Xn⊥ , kx∗n k = 1, and for
each x ∈ C we have that
hx∗n , xi = hPn∗ (zn∗ ), xi/kPn∗ (zn∗ )k = hzn∗ , Pn (x)i/kPn∗ (zn∗ )k
−1
−cn
≥ n+1 ∗ ∗ ≥ n+1 ,
4
kPn (zn )k
4
since Pn (x) ∈ Cn . Now choose an integer kn ≥ max{kn−1 , n + 1} and real numbers
Pkn
αi , i = 1, . . . , kn , so that for vn =
i=1 αi xi , it holds that kvn k ≤ 3/2 and
hx∗n , vn i ≥ 1. Since the spaces Xn are increasing, their union is dense in X, and
follows that the sequence {x∗n } converges to zero in the
each x∗n annihilates Xn , it P
∞
∗
weak topology. Put v = i=1 4−i vi . Then hx∗n , x + vi ≥ 0 for each x ∈ C and
each n ∈ N . Indeed, if x ∈ C, then
hx∗n , x + vi = hx∗n , xi + hx∗n ,
n−1
X
∞
X
4−i vi i + hx∗n , 4−n vn i + hx∗n ,
i=1
≥ −4−(n+1) + 0 + 4−n − 4−n
4−i vi i
i=n+1
∞
X
4−i kvn+i k ≥ 0,
i=1
where the last inequality follows from the fact that kvi k ≤ 3/2. Since X is separable
and x∗n is a weak∗ -null sequence in the sphere of X ∗ , it has a subsequence which
is weak∗ basic (cf. [LT], pg. 11). Without loss of generality, suppose that {x∗n } is
already such a sequence. Let {x∗n }⊥ be the annihilator of {x∗n } in X. Then (again,
see [LT]), the quotient space Y = X/{x∗n }⊥ has a basis {yn } and x∗n = T ∗ (yn∗ ),
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
1796
EVA MATOUŠKOVÁ
where T : X → Y is the quotient map and {yn∗ } are the coeficient functionals of
{yn }. Then,
hyn∗ , T (x + v)i = hT ∗ (yn∗ ), x + vi = hx∗n , x + vi ≥ 0
for all x ∈ C and n ∈ N . Consequently, the projection of C is contained in the
positive cone of Y .
The construction of the measure on `1 in the proof of the following lemma is
a modification of the proof in [BN] (p. 54), where it is shown that for p ≥ 1 the
positive cone of `p is Haar null.
Lemma 2. Let X be a Banach space with a normalized Schauder basis {xn }, and
suppose there exist p ≥ 1 and K > 0 so that
n
n
X
X
1
|βi |p ) p ≤ Kk
βi xi k
(
(1)
i=1
i=1
for all real numbers βi , i = 1, . . . , n and n ∈ N . Then the positive cone Q of X is
Haar null.
Proof. Firstly, let us construct a probability measure µ on the positive cone of `1
such that whenever (βn ) is a sequence in `p then
µ{(αn ) ∈ `1 : αn ≤ βn } = 0.
To this end, choose a point (an ) in `p+1 \ `p with each an > 0; for example put
an = n−1/p . Fix independent random variables fn : [0, 1] → [0, 1] such that
λ{t ∈ [0, 1] : fn (t) = 0} = 1 − apn
λ{t ∈ [0, 1] : fn (t) = an } =
and
apn ,
where λ denotes Lebesgue measure. Let F (t) = (fn (t)) be the mapping from [0, 1]
to R∞ , the countable product of the reals with the product topology. Then F
is Borel measurable because it is coordinatewise measurable, and we can define a
Borel probability measure νo on R∞ as follows:
νo (A) = λ{t ∈ [0, 1] : F (t) ∈ A},
for every Borel subset A of R∞ . Let ν be the completion of νo . Consider the natural
imbedding g((αn )) = (αn ) of `1 into R∞ . Since `1 is a complete metric space and
g is continuous, g carries Borel subsets of `1 to analytic subsets of R∞ , and the
latter, by a theorem of Szpilrajn-Marczewski, are ν-measurable since ν is complete.
∞
(Or, more easily,
observe that closed
T
Pn balls in `1 are closed subsets of R ; namely,
∞
B `1 (0, r) = n∈N {(xi ) ∈ R :
i=1 |xi | ≤ r}. Therefore, Borel subsets of `1 are
Borel also in R∞ , since `1 is separable.) Hence, if we define µ(B) = ν(g(B)) for
Borel subsets of `1 , then µ is a Borel measure on `1 . To show that µ is a probability
measure on `1 it is enough to check that (fn (t)) ∈ `1 for almost all t ∈ [0, 1]. This
follows because we have
Z 1X
∞ Z 1
∞
∞
X
X
|fn (t)|dt =
|fn (t)|dt =
ap+1
< ∞,
n
0 n=1
P
n=1
0
n=1
and this proves that
|fn | < ∞ almost everywhere. Hence µ(`1 ) = 1. Fix (βn ) ∈
`p . If βn < 0 for some n, then, clearly, µ{(αn ) ∈ `1 : αn ≤ βn } = 0; otherwise,
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
HAAR NULL SETS
1797
define
θ = µ{(αn ) ∈ `1 : αn ≤ βn } = λ{t ∈ [0, 1] : fn (t) ≤ βn for all n}
=
∞
Y
λ{t ∈ [0, 1] : fn (t) ≤ βn }
n=1
=
Clearly, θ = 0 if η =
Y
(1 − apn ).
βn <an
P
βn <an
∞=
apn = ∞. Since (an ) ∈
/ `p , if η < ∞ then
X
βn ≥an
apn ≤
X
βn ≥an
βnp ≤
∞
X
βnp < ∞
n=1
P∞
and we have a contradiction. Now, let T ((αn )) = n=1 αn xn be the natural linear
imbedding of `1 into X. Then the formula τ (B) = µ(T −1 (B)) for Borel subsets B
of X defines a Borel probability measure on X. Let us P
show that τ (x − Q) = 0 for
∞
every translate x − Q of the negative cone −Q. If x = n=1 βn xn , then
τ (x − Q) = µ{(αn ) ∈ `1 : αn ≤ βn } = 0,
P∞
where the latter equality follows from the fact that if n=1 βn xn ∈ X, then (βn ) ∈
`p by (1). Hence, the negative cone −Q is Haar null in X, and consequently so is
Q.
We remark that the convexity assumption in the following lemma is essential; if
S is the unit sphere of a Banach space X and P a projection onto a subspace of X
then the relative interior of P (S) is nonempty. The lemma holds also, by a result of
Vašák [V] (see [S] for a different proof), for weakly countably determined Banach
spaces. The proof of the lemma was suggested to us by C. Stegall.
Lemma 3. Let X be a weakly compactly generated Banach space and C a closed
and convex subset of X with empty interior. Then there exists a projection P onto
a separable subspace Z of X such that the interior of P (C) in Z is empty.
Proof. Since X is weakly compactly generated, by the theorem of Amir and Lindenstrauss [AL], there exists a sequence of projections Pn : X → X so that Pn (X)
is separable for all n ∈ N , Pm Pn = Pn Pm = Pmin{m,n} for all m, n ∈ N , and
P (x) = lim Pn (x) for x ∈ X exists and is a projection onto ∪Pn (X). Moreover, one
can construct these projections so that if Pn−1 is already constructed and Dn ⊂ X,
An ⊂ X ∗ are given separable sets then Pn can be chosen so that Dn ⊂ Pn (X) and
An ⊂ Pn∗ (X ∗ ) (see [AL]). Let us construct by induction a sequence of such projections as follows. Let Po be any projection of X onto a separable subspace. If Pn−1
was already constructed, choose a countable set Dn ∈ X \C so that Pn−1 (C) ⊂ Dn .
This is possible, because the interior of C is empty and Pn−1 (C) is separable. Since
C is closed and convex, each point of Dn can be strictly separated from C by some
x∗ ∈ X ∗ , kx∗ k = 1. Therefore we can choose a countable set An in the unit ball of
X ∗ so that for each x ∈ Dn there exists some x∗ ∈ An and δ > 0 so that
(2)
hx∗ , x − zi ≥ δ for all z ∈ C.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
1798
EVA MATOUŠKOVÁ
Choose Pn as above so that Dn ⊂ Pn (X) and An ⊂ Pn∗ (X ∗ ). From the continuity
of the projections it follows, that
P (C) ⊂
∞
[
Pn (C) ⊂
n=0
∞
[
Dn .
n=1
S
S
S
Also, P (x) = x and P ∗ (x∗ ) = x∗Sfor all x ∈ Dn and x∗ ∈ An . Let x ∈ Dn
be given; choose δ > 0 and x∗ ∈ An so that (2) holds. Then,
dist(x, P (C)) = dist(x, P (C)) ≥ inf hx∗ , x − P (z)i
z∈C
∗
= inf hx , P (x) − P (z)i = inf hP ∗ (x∗ ), x − zi
z∈C
= inf hx∗ , x − zi ≥ δ > 0.
z∈C
S
Since Dn contains P (C) and each point of
the interior of P (C) is empty.
S
z∈C
Dn has positive distance to P (C),
The proof of the following theorem shows that if X is a reflexive Banach space
with the property that the positive cone of each quotient of X with Schauder basis
is Haar null then for every closed and convex C ⊂ X with empty interior there
exists a Radon probability measure µ on X so that µ(C + x) = 0 for all x ∈ X.
Superreflexive spaces have this property, since basic sequences in superreflexive
spaces satisfy estimates similar to (1) in Lemma 2. Namely, a Banach space is
superreflexive if and only if 0 < 2k < ε ≤ 1 < K; then there exist p, q > 1 such
that for any normalized basic sequence {xn } in X with characteristic not less then
ε we have
n
n
n
X
X
X
1
1
p p
k(
|βi | ) ≤ k
βi xi k ≤ K(
|βi |q ) q
i=1
i=1
i=1
for any real numbers β1 , . . . , βn (see [GG], or [J]).
Theorem 4. Let X be a superreflexive Banach space and C ⊂ X a closed and
convex set with empty interior. Then there exists a Radon probability measure µ
on X so that µ(C + x) = 0 for all x ∈ X. In particular, in separable superreflexive
Banach spaces, closed and convex sets with empty interior are Haar null.
Proof. As in the previous lemma, let P be a projection of X onto a separable
subspace Z of X such that the interior of P (C) in Z is empty. We will show
that P (C) is Haar null in Z. Then, if µ is a Borel probability measure on Z for
which µ(P (C) + z) = 0 whenever z ∈ Z we can extend µ on Borel subsets of
X by the formula µ(A) = µ(A ∩ Z). Then µ has the required properties since
(C + x) ∩ Z ⊂ P (C + x) ⊂ P (C) + P (x) for each x ∈ X. Since a countable union of
Haar null sets is Haar null, we can suppose that D = P (C) is bounded. If D is not
spanning in Z, there exists z ∈ Z so that the intersection of any translate of the line
segment [0, z] and D contains at most one point. Consequently, any translate of D
is a null set for the Lebesgue measure sitting on [0, z]. Suppose that D is spanning.
By Lemma 1 there exists a quotient space Y = T (Z) of Z with a basis {yn } and
v ∈ Z so that T (D + v) is contained in the positive cone Q of Y , where T is the
quotient mapping. By considering {yn /kyn k} instead of {yn }, we can suppose that
the basis of Y is normalized. Since X is superreflexive, Y is also superreflexive,
and consequently the basis admits an estimate of type (1). Hence the positive
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
HAAR NULL SETS
1799
cone Q of Y is by Lemma 2 Haar null; denote by ν a probability measure on Y ,
for which ν(y + Q) = 0 for all y ∈ Y . Since T is a surjective continuous linear
operator, the inverse of T admits, by a theorem of Bartle and Graves (see [BP]),
a continuous selection f . Consequently, the formula τ (A) = ν(f −1 (A)) for Borel
subsets A of Z defines a probability measure on Z. Let z ∈ Z be given. Then
f −1 (D + z) ⊂ T (D + v) + T (z − v) ⊂ Q + T (z − v), and consequently
τ (D + z) = ν(f −1 (D + z)) ≤ ν(T (D + v) + T (z − v))
≤ ν(Q + T (z − v)) = 0.
Hence τ (D + z) = 0 for each z ∈ Z, and D is a Haar null set.
We wish to thank L. Zajı́ček for directing our attention to this problem, and C.
Stegall for help with Lemma 3.
References
[A] N. Aronszajn, Differentiability of Lipschitzian mappings between Banach spaces, Studia
Math. LVII (1976), 147-190. MR 54:13562
[AL] D. Amir and J. Lindenstrauss, The structure of weakly compact sets in Banach spaces,
Ann. of Math. 88 (1968), 35-46. MR 37:4562
[BP] C. Bessaga and A. Pelczynski, Selected topics in infinite-dimensional topology, PWN,
Warszawa 1975. MR 57:17657
[BN] J.M. Borwein and D. Noll, Second order differentiability of convex functions in Banach
spaces, Trans. A.M.S. 342 (1994), 43-81. MR 94e:46076
[C] J.P.R. Christensen, Topology and Borel structure, North-Holland, Amsterdam 1974. MR
50:1221
[GG] V.I. Gurarii and N.I. Gurarri, On basis in uniformly convex and uniformly smooth spaces,
Izv. Akad. Nauk SSSR 35 (1971), 210-215. MR 44:780
[J] R.C. James, Super-reflexive spaces with bases, Pacific J. of Math., Vol. 41, 2 (1972), 409-419
MR 46:7866
[LT] J. Lindenstarauss and L. Tzafriri, Classical Banach spaces I., Springer-Verlag, Berlin 1977.
MR 58:17766
[MS] E. Matoušková and C. Stegall, A characterization of reflexive Banach spaces, to appear in
Proc. A.M.S. MR 96g:46005
[MZ] E. Matoušková and L. Zajı́ček, Second order differentiability of integral functionals, to
appear in Czechoslov. Math. Journ.
[PT] D. Preiss and J. Tišer, Two unexpected examples concerning differentiability of Lipschitz
functions on Banach spaces, to appear in Proceedings of GAFA seminar, Tel Aviv.
[S] C. Stegall, A proof of the theorem of Amir and Lindenstrauss, Israel Jour. of Math., 68
(1989), 185-192. MR 91b:46017
[V] L. Vašák, On one generalization of weakly compactly generated Banach spaces, Studia Math.
70 (1981), 11-19. MR 83h:46028
Department of Mathematical Analysis, Charles University, Sokolovská 83 , 18600
Prague, Czech Republic
Institut für Mathematik, Johannes Kepler Universität, Altenbergerstraße, A-4040
Linz, Austria
E-mail address: [email protected]
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use